Research map / Architectures and efficiency / Quantization
TRQ: Ternary Neural Networks With Residual Quantization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ternary stem-residual quantization instead of direct thresholding | ResNet-18 32-32-64-128 · CIFAR-10 | accuracy: 87.5 -> 89.3 | Helped |
| Use ternary stem-residual quantization instead of direct thresholding | VGG-Small · CIFAR-100 | accuracy: 61.8 -> 65.4 | Helped |
| Use ternary stem-residual quantization instead of direct thresholding | ResNet-18 · ImageNet | Top-1 accuracy: 61.6 -> 62.6 | Helped |
| Use ternary stem-residual quantization instead of direct thresholding | ResNet-34 · ImageNet | Top-1 accuracy: 65.2 -> 66.2 | Helped |
| Add multi-batchnormalization to TRQ | TRQ ResNet-18 · ImageNet | Top-1 accuracy: 62.6 -> 64.4 | Helped |
| Add multi-batchnormalization to TRQ | TRQ ResNet-34 · ImageNet | Top-1 accuracy: 66.2 -> 68.2 | Helped |
| Apply residual quantization to activations as well as weights | TRQ ResNet-18 · ImageNet | Top-1 accuracy: 64.5 -> 65.7 | Helped |
| Add a learnable coefficient α rather than omit it (ablation) | TRQ ResNet-18 16-16-32-64 · CIFAR-100 | accuracy: 52.1 -> 54.9 | Helped |
| Optimize α rather than fixing it at 0.6 (ablation) | TRQ ResNet-18 32-32-64-128 · CIFAR-100 | accuracy: 61.3 -> 62.7 | Helped |
| Optimize α rather than fixing it at 0.6 (ablation) | TRQ VGG-Small · CIFAR-100 | accuracy: 60.5 -> 65.4 | Helped |
| Increase residual quantization bit width from 2bit to 4bit (ablation) | TRQ ResNet-18 · CIFAR-100 | accuracy: 56.2% → 58.3% → 58.5% | Helped |
| Use residual quantization rather than the multi-bit baseline (ablation) | TRQ ResNet-18 · CIFAR-100 | accuracy: 0.4%, 1.1%, 1.0% | Helped |
Closest papers by meaning
- Trained Ternary Quantization (ICLR 2017 · similarity 0.92)
- RTN: Reparameterized Ternary Network (AAAI 2020 · similarity 0.88)
- HitNet: Hybrid Ternary Recurrent Neural Network (NeurIPS 2018 · similarity 0.87)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.85)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.85)
- AffineQuant: Affine Transformation Quantization for Large Language Models (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.